Datasets:
Register interest in prospective SportVISTA research access
The current v0.1.0 release is documentation-only. This form records prospective, project-specific research requests; submission does not imply that Dataset Materials are currently available or that access will be granted.
SportVISTA v0.1.0 distributes no raw media or row-level metadata. Any future access would be personal, project-specific, time-limited, revocable, limited to cleared materials, and governed by the SportVISTA Research-Only Dataset License. Submission does not authorize redistribution or publication of media excerpts.
Log in or Sign Up to review the conditions and access this dataset content.
SportVISTA
SportVISTA documents an internal video corpus collected from club- and organizer-operated sports streams under heterogeneous operational conditions. The archive contains fixed, tripod-mounted, handheld, and automated-camera streams from indoor and outdoor venues, with substantial variation in duration, viewpoint, framing, resolution, frame rate, bitrate, and production quality. AAC is reported in the container metadata, but this observation does not by itself establish successful decoding, non-silent content, or audio-video synchronization for every file.
The frozen paper snapshot contains 8,144 indexed MP4 files, approximately
4,285 file-hours, and 3.04 TB. It covers ten named sports and the
auxiliary operational-negative class other. File counts and file-hours do not
represent unique sporting-event counts or durations because the archive may
contain interrupted streams and multiple renditions of one event.
Controlled-release status. As of August 20, 2026, the public
v0.1.0release is documentation-only. It contains public documentation, governance terms, aggregate statistics, schema information, contribution guidance, and this prospective access-request form. It does not distribute raw media, row-level metadata, benchmark manifests, or other Dataset Materials. Submitting the form does not imply that a data release is currently available, that an application can presently be approved, or that access will be granted.
Dataset scope
The ten named sports are badminton, basketball, 3×3 basketball, football, handball, rugby, squash, table tennis, tennis, and volleyball.
The other class is not a sport. It contains club-originated test captures,
interrupted captures, and recordings that are too short or insufficiently
informative to support assignment to a named sport. It is an auxiliary
operational-negative cohort, not a complete open-set-recognition benchmark.
Provenance and labels
The indexed files originate from operational live streams submitted by clubs and event organizers to the former French sports-streaming platform CHAMP'S. According to retained platform records, participating clubs completed a partner-registration form, declared the sport associated with their streams, and accepted the applicable participation conditions. These records describe the collection regime but do not establish research-redistribution rights or a verified competition-level annotation for every indexed file.
Sport labels originate from the sport declared directly by the club when
configuring a stream. They were normalized to the ten-class SportVISTA taxonomy.
These are source-provided, recording-level labels rather than dense frame- or
event-level annotations. Platform tests and recordings without sufficiently
reliable evidence for a named sport are assigned to other.
Original public live availability is not treated as authorization for research redistribution. Any future controlled delivery would be limited to material cleared for the approved request.
Planned research uses
Subject to legal and data-protection review and to the contents of a future cleared release, SportVISTA may support:
- video-only sport classification;
- supervised rejection using the documented operational-negative cohort;
- representation learning on heterogeneous long-form sports video;
- robustness and data-quality analysis across technical metadata slices; and
- active annotation and versioned community label layers.
Audio-only, cross-modal, and other audio-visual tasks require a manifest-defined subset with successful decoding, non-silent content checks, and verified temporal alignment. Dense event spotting, player or ball tracking, pose estimation, tactical analysis, captioning, and identity-related tasks are not part of the current documentation layer.
Prospective controlled access
The form on this page records prospective, project-specific requests. No
row-level index or media release is currently available through v0.1.0.
If and when a cleared release becomes available, requests would be reviewed per person and per project, considering the applicant's identity, affiliation, research purpose, requested sports, requested release layers, requested volume, and acceptance of the applicable research-use terms.
Any future approval would:
- be limited to non-commercial academic or scientific research;
- apply only to the approved person, project, release, and subset;
- not automatically extend to colleagues or an institution;
- be time-limited and revocable;
- be reduced to the minimum volume necessary for the approved project; and
- remain subject to removals and takedown-driven updates.
A future metadata release requires legal and data-protection review. Any subsequent controlled-media release additionally requires file-level rights clearance, decoder validation, and audio-video synchronization checks. No calendar date is claimed before those conditions are satisfied.
Prohibited uses
SportVISTA Dataset Materials must not be used for:
- commercial activities or for-profit benefit without a separate written agreement;
- training, fine-tuning, evaluating, or improving a general-purpose, foundation, or generative model;
- redistribution, mirroring, sublicensing, credential sharing, or transfer to an unapproved person;
- face recognition, biometric identification, person re-identification, profiling, surveillance, or sensitive-attribute inference;
- contacting, targeting, harassing, or attempting to identify recorded participants; or
- decisions affecting employment, education, insurance, credit, policing, eligibility, or access to services.
Privacy and responsible research
Recordings may contain identifiable athletes, staff, spectators, signage, voices, crowd audio, and young participants. Researchers must minimize access, protect Dataset Materials, obtain applicable ethics approvals, and report suspected rights, privacy, security, or integrity issues before public disclosure.
Publishing a frame, clip, audio excerpt, or other media sample would require prior written authorization. Standard dataset access would not grant publication or redistribution rights.
Data characteristics and known limitations
- The distribution is strongly long-tailed; football represents 62.5% of the frozen snapshot by indexed-file count.
- Multiple spatial renditions of one capture are separate MP4 files but not independent sporting events.
- Some files are short, interrupted, technically atypical, or available at only a subset of resolutions.
- The index records 41 files with 0 x 0 resolution metadata, 41 with unavailable video-codec metadata, and 23 with a nominal 90,000-FPS time-base artifact.
- Clubs, venues, channels, devices, and production graphics can introduce correlated nuisance factors and leakage risks.
- Overlay prevalence has not been quantified, and no dedicated forensic audit for invisible watermarks has been performed.
- Container-level AAC metadata does not establish successful audio decoding, non-silent content, or audio-video synchronization for every file.
- Sport labels are not dense event, identity, pose, tracking, or tactical annotations.
- No systematic manual label audit or inter-annotator agreement study is reported.
- Competition status is not audited file by file, and geographic and demographic coverage is incomplete.
Researchers must report the exact release, cohort, task profile, split, sampling strategy, modality, exclusions, grouping rules, and safeguards used. Channel-aware and event-aware splits are recommended whenever identifiers are available.
Versioning and reproducibility
Any future released artifacts must use immutable media identifiers and explicit version numbers. Corrections, additions, removals, metadata repairs, and takedowns require a new release and changelog entry; an existing corpus view must never be silently overwritten.
Benchmark manifests should record:
- the source index and label-layer versions;
- retained media identifiers and ordered filters;
- grouping keys and split seeds;
- clip-sampling and preprocessing rules;
- decoder and library versions;
- exclusions with machine-readable reasons;
- checksums; and
- prediction logs and per-class confusion matrices when applicable.
Community contributions
Discussions and pull requests are enabled on this project page. Researchers may propose annotation layers, metadata corrections, schemas, evaluation scripts, or documentation improvements. Contributions must reference immutable media identifiers, document their methodology and provenance, and preserve contributor authorship. Accepted contributions would be reviewed independently and released as separate versioned layers after a distributable base identifier registry becomes available.
Citation
Lazhar Bouacha, Laurent Wendling, and Camille Kurtz. 2026. "SportVISTA: Scaling Multimodal Audio-Visual Representation Learning to Amateur and Semi-Professional Sports Videos." In 9th Int. Workshop on Multimedia Content Analysis in Sports (MMSports '26). ACM. https://doi.org/10.1145/3841455.3841528
The exact SportVISTA documentation or dataset release used must always be reported.
Maintainer
SportVISTA is maintained by Lazhar Bouacha as part of the SportIQ research project.
- Downloads last month
- 84